Cognition SWE-2 Challenges Frontier Coding Cost Curves

Cognition's SWE-2, a Kimi K3 post-trained coding model, reportedly matches Claude Fable 5.1 on FrontierCode at 64% lower cost and supports selectable effort levels with fewer turns. Vendor-run benchmarks, poor Terminal-Bench 4 performance, lack of weights, no general API, and weaker open-weight repository results leave portability and generality unresolved; self-hosting remains attractive mainly for control, privacy, and cost experimentation.